Machine Learning for the Deterioration of Patients on Hospital Wards

Early Warning Score Systems assign scores to a patient’s vital signs to detect adverse events in order to secure timely medical response. However, most systems do not account for variations in vital signs due to aging changes or sex differences, which affects the accuracy of current systems for different deteriorating patient sub-groups. We aimed to develop an Ageand Sex-Based Early Warning Score derived from statistical distributions of vital signs across ages and sexes. The data is obtained from four hospitals in Oxford, United Kingdom, which contains a total of 48,433 patient admissions to the wards with documented sex and age, including 3,182 patient admissions with at least one adverse event. Vital sign distributions were compared at each year of age (age, 16 yr to 99 yr) as well as sexes (females and males) using aggregated cumulative distribution functions. The centiles of the vital sign distributions modeled the normal ranges of the Ageand Sexbased Early Warning Score. An adverse event was defined as the occurrence of in-hospital mortality, cardiac arrest or ICU admission within the first 24 hours of stay. The area under the receiver operating curve for the Ageand Sex-Based Early Warning Score was higher than that of existing systems, namely Centile-based Early Warning Score (0.828 vs 0.730; p <0.001), the National Early Warning Score (0.828 vs 0.808; p <0.001) and the Modified Centile-based Early Warning Score (0.828 vs. 0.825; p <0.001). The Ageand Sex-Based Early Warning Score also outperformed the current systems over the nonelderly patients (age, 16 yr to 40 yr), and amongst both sexes, females and males. Our results suggest that accounting for ageand sex-related vital sign changes more can accurately detect deterioration of non-elderly patients prior to an adverse event than current methods. Further investigations of ageand sexspecific methods are necessary to decrease preventable deaths.